Divya van Mahajan

Point-Counterpoint: Consumers Will Never Pay for AI

ai-daily-brief-podcast
Listen to episode →

Source lookup was blocked by sandbox permissions, so the Overview records that no URL was supplied rather than guessing one.

Overview

This is an episode of The AI Daily Brief, a daily news podcast and video about developments and debates in AI. The host is not named in the transcript, and no channel or source URL was provided with it — the episode slug is 2026-10-06-point-counterpoint-consumers-will-never-pay-for-ai.

Source video: no source URL was provided with the transcript.

The episode has two segments. It opens with a “Headlines Edition” (roughly five minutes of news items), then moves to the main episode, whose central question is the one the title names: given that only ~2.2% of US households pay for an AI subscription, does the remaining 98% represent a vast untapped market, or a population that will never pay for AI subscriptions — and therefore requires different AI products, or at least different business models, to matter to the large AI companies?

The main episode’s starting point is the seventh edition of Andreessen Horowitz’s “Top 100 Consumer AI Apps” report, which for the first time includes a revenue-ranked list alongside the web-traffic and mobile-usage lists. The host frames the discussion as a genuine point-counterpoint, presenting both the “we’re just early” and the “there’s no utility here for normal people” positions, and argues that even if the pessimists are right about the average consumer, two factors — the uncapped spend of power users, and non-subscription monetisation such as advertising — mean the industry’s financial thesis does not necessarily collapse.

Prerequisites

To follow the episode, a reader should be familiar with:

  • Consumer software metrics — monthly active users (MAU), unique monthly visits, monthly revenue, net new subscriptions, ARR, and why ranking the same market by each produces different leaderboards.
  • Open-weight vs. closed models, model parameter counts, and the idea of using a cheap, efficient base model for further post-training.
  • AI coding/agentic benchmarks and the practice of “benchmark maxing” — optimising for a published benchmark rather than general capability.
  • Reinforcement learning as a post-training stage, and the notion of distributed RL infrastructure.
  • Agentic AI concepts: agents that drive a browser, read email, or call APIs on a user’s behalf; “agent swarms” and bulk API scraping.
  • Enterprise SaaS economics: customer concentration risk, TAM (total addressable market), per-seat subscription price ceilings, and why IPO disclosures invite scrutiny of both.
  • Technology adoption curves — the historical time for a technology to move from ~2% to ~50% of US households.
  • Advertising as the default monetisation model of the consumer internet.

Main Points

Segment 1 — Headlines

Reflection AI releases Beam, a US-origin open-weight model

  • Following the prior day’s episode about a rumoured American open-weight model, Reflection AI unveiled Beam, positioned as the US-trained challenger to the Chinese open models competing for token share.
  • Size: 501 billion parameters — roughly comparable to NVIDIA’s NemoTron 3 Ultra, smaller than GLM 5.3, and small next to Kimi K3’s 2.8 trillion parameters. Availability is limited to early testers, so assessment rests on reported benchmarks.
  • Reported scores: 44.4 on the coding benchmark the transcript renders as “DeepSwee” (vs. GLM 5.2 at 44, and QEN 3.8 Max at 51); 80.1 on Terminal Bench 2.1 for agentic coding (vs. NemoTron Ultra 56.4 and “Inkling” 63.8, but slightly behind GLM 5.2 and six points behind QEN 3.8 Max). Across Reflection’s chosen benchmark set, Beam led only on SWE-Bench Verified — a chart that included no Chinese model scores.
  • Benchmark selection noted as telling: Reflection did not benchmark on Terminal Bench 4.0 (which has exposed significant benchmark maxing since its release the prior month), and did not compare Beam against Chinese leaders GLM 5.3, Kimi K3, or DeepSeek 4.1 Flash.
  • Efficiency is the pitch: Reflection claims three-to-four times better inference-compute efficiency than GLM 5.2, which would make Beam a cost-effective base for enterprise post-training — the use case the show has described as increasingly attractive to enterprises, and evidently central to Reflection’s strategy. The model was pre-trained from scratch and then put through what Reflection believes is the largest publicly documented RL run ever, using distributed RL infrastructure — a method the host calls wildly underexplored for open models.
  • Reception was split. Artificial Analysis had not finished independent testing but offered directional support: “Early indicators suggest Beam will be one of the most token-efficient open models we’ve seen for its level of intelligence.” Kyle Chan of the Brookings Institution called it impressive but “still behind GLM 5.3 and Kimi K3,” noting the real game-changer would be an American open-source model on par with the best Chinese models. Nathan Lambert congratulated the team on shipping a first model; former White House advisor Sriram Krishnan called leading US open models critical to winning the AI race.
  • The host’s own reading: the sense of letdown was manufactured by two days of leaks and hype. Absent that, the reaction would have resembled scoring the first goal while two down — not proof of a comeback, but a marked improvement on the position minutes earlier.

New York City Council holds a hearing on existential risk

  • The hearing featured testimony from prominent AI-safety figures, including former Anthropic employee Jacob Coxon, who had gone viral weeks earlier. The host reports nothing substantively new was discussed; the hearing largely served as a canvas for AI regulations proposed at the New York State level.
  • Council Speaker Julie Menon opened by saying “the idea that artificial intelligence is going to self-regulate defies all reason,” and suggested New York’s bills could serve as a national model.
  • New York’s existing and pending measures: a fairly robust set of AI regulations already in place with tougher consumer protections phasing in over coming years; proposals for an AI kill switch and whistleblower protections before the state legislature; and a one-year moratorium on data-centre construction imposed by Governor Kathy Hochul.
  • The Speaker referenced Sam Altman’s recent remark that the world should “accept some bad things happening for the benefits of this technology,” and pressed AI-company representatives to attach a probability to extinction risk. None gave a number.
  • OpenAI’s submission reframed the agenda, centring present-day safety issues rather than hypothetical recursive self-improvement; in a letter to the Council it offered broader collaboration with the city’s cyber defenders and noted existing work with the state.

An “agent fleet” detected on Chinese infrastructure

  • A group of self-styled “agent swarm chasers” published a preliminary report on Sunday describing several thousand agents running on Tencent’s AI infrastructure and sending bulk requests to Alibaba’s map service, AMAP.
  • The researchers deliberately avoided the term agent swarm, calling it an agent fleet instead: they found no evidence of coordination between the agents, unlike OpenAI’s agents during the Hugging Face incident.
  • Intent appears benign or at most rule-breaking — the agents requested information about local landmarks such as parks, zoos, and hospitals; at worst they circumvented Alibaba’s API rules to bulk-scrape data.
  • This may be the first documented sign of large-scale agent fleets originating from Chinese servers, though the host cautions that agentic swarm chasing is new enough that this could simply be the first documented chase.
  • The host’s view: this kind of research — which drew a like-minded group to a hackathon the prior weekend — will be an inevitable part of resiliency infrastructure, and tracking fleets publicly helps demystify what they are actually doing as they crawl the internet.

Microsoft and Meta cut their Claude spending ahead of Anthropic’s IPO

(Note: the transcript’s automatic transcription renders “Claude” as “cloud” throughout this item; the content — Anthropic forecasts, Copilot, Claude Code, MetaCode — makes the intended word unambiguous.)

  • Microsoft, per reporting from The Information, has cut its projected Claude bill by a third; sources said it was on track to spend $1 billion on internal use earlier in the year. Programmers were asked to ration usage as a cost-saving measure and to switch to Microsoft’s in-house models. This figure covers internal use only, not spending by Microsoft’s customers.
  • Microsoft’s separate forecast of $2 billion in Anthropic spend to power Copilot features has not changed. It has substituted in-house models and OpenAI’s GPT-5.6 for many queries, but sources said Copilot customer growth has offset the effect of switching away from Claude.
  • Meta has cut usage in half, having dramatically reined it in during the spring. At peak, roughly 60,000 of Meta’s then-78,000 employees were using Claude Code; after 10% layoffs and in-house model releases, that is down to 30,000 — a drop almost exactly mirrored by adoption of its in-house tool MetaCode, which now also has 30,000 users. Meta was on pace to spend several billion with Anthropic as of June; it now spends around $105 million per month.
  • Why the story matters now: Anthropic’s upcoming IPO. Wall Street is digging into the business, and leaked financial disclosures — mostly 2025 figures of what the host calls limited relevance to the company today — included one that drew attention: two customers accounted for a quarter of Anthropic’s 2025 revenue, speculated to be some combination of Meta, Microsoft, and Cursor. Anthropic has told investors concentration has since broadened: 6,000 customers with annual bills above $100,000, but only about 100 spending over $10 million annually — so belt-tightening at the top remains material.
  • The host’s caution: Microsoft and Meta are not representative of the broader customer base. They are competing labs that used Anthropic models begrudgingly because those models were so much better than their in-house options; as that gap closes — or as the floor of in-house models rises — the switch becomes justifiable. That logic does not apply to the average customer.
  • Two framings of customer concentration, which the host expects to define the pre-IPO debate: purely as risk (a handful of customers cutting spend reshapes revenue), or as evidence of early adoption and therefore growth headroom across the rest of enterprise. As with most binaries, the host says, it is probably some of both.

Segment 2 — The main episode

The a16z Top 100 Consumer AI Apps, seventh edition

  • Two data sources underpin the report. Consumer traffic data — the one series comparable across all seven editions — yields a top 50 by unique monthly web visits and a top 50 by mobile monthly active users. For the first time, a16z also used data from YipitData to publish a top 50 by monthly revenue, and in many cases that list tells a very different story.
  • Fewest debuts ever: only 11 products appeared on an a16z list for the first time — the smallest cohort across seven editions. The host reads this as at least inching toward a consolidation phase, where instead of a thousand flowers blooming, patterns and network effects are taking hold within categories.
  • Consumer AI’s first real market is work. a16z argues people pay for tools that help them build, create, and get things done. Playful or companion-style AI appears on the mobile and web charts, but the revenue chart is dominated by tools used in some professional capacity.
  • At the top: among leading LLMs — ChatGPT, Gemini, and Claude — relative positions have stayed fairly consistent among consumers. Over 2026, Claude caught up to Gemini on paid US subscribers, though both remain far behind ChatGPT. Claude converts users to higher-cost tiers better. For a large part of the year Claude led ChatGPT on net new subscriptions, but that flipped in July and August, with ChatGPT significantly outperforming.

Three ways startups still differentiate against the incumbents

  • Own a differentiated model — clearest in creative products: music generation platform Suno and voice platform ElevenLabs.
  • Own a multi-model experience — Cursor, which supports models from several labs (the host notes this may change as it is absorbed further into SpaceX), and OpenRouter, built specifically to give developers access to multiple providers from one platform.
  • Extreme focus on a specific audience need — Open Evidence for physicians, Venice for privacy-focused users.

What surprised the community

  • A common reaction was simple unfamiliarity. Peter James: “There are so many surprises here to me. Superhuman at number four? Suno at number seven? Open art above Notion and Figma? Also, WTF is Plaud? I’m absolutely bamboozled.” The host attributes this to bubbles in which whole usage categories escape notice — and offers Plaud as his own example: he had tracked it only for the show until an interior designer at his house pulled one out, because the walk-around-the-house form factor suits how they actually work.
  • Suno at number seven on revenue is widely misread: the assumption is that paying users are professional or aspiring musicians, when a large share of usage is entertainment — people making fun, ridiculous songs for family and friends, and enjoying it enough to pay.
  • The show’s running themes appear on the charts: agentic coding platforms Replit and Lovable near the top of the revenue list; OpenRouter at 13 even on a consumer chart, indicating how much people seek cost optimisation; and Noose, Genspark, and Manus all charting, reinforcing growing competition in personal AI assistants.
  • Stickiness surprised people too. Perplexity has not been a major topic in enfranchised AI circles for some time, yet sits at number 12 on unique monthly visits, 25 on mobile MAU, and 9 on monthly revenue. n8n — in many ways a precursor to the agentic era — remains on the revenue charts, which the host reads as evidence of how slowly behaviour changes once people have invested in a platform, especially if they did the work to automate a key workflow.

The dominant finding: power users versus everyone else

  • The top 1% of AI spenders now outspend the bottom 50% combined, averaging $903 a month against a median paying customer’s $25. Crucially, this is the top 1% of people who spend anything — itself a vanishingly small share of all AI users.
  • What the 1% buy disproportionately is tools to build and automate, tools to create content, and tools for work. Their purchase rate for Notion is over 9× the average; for video platforms Runway and Higgsfield, 10× and 15× respectively; and for n8n, the top 1% are roughly 23× as likely to be paying as all other buyers.

The 98%: untapped market or hard ceiling?

  • The chart most often paired with the 1% concentration data comes from a16z’s separate State of Markets report: as of April of this year, per the research firm the transcript names as PNC Research, only 2.2% of US households had a paid AI subscription. The question is whether 98% of households aren’t paying yet, or aren’t ever going to.
  • The optimistic case — we’re early. Adam at RepView: “I think this is just a reflection on early adoption and how much TAM there actually is as mass market adoption proliferates.” Ruben Hasid offered historical comparisons for going from 2% to 50% US adoption: 19 years for the computer, 15 for cars, 6.5 for the smartphone. Others compared AI to existing household spend — Co-founders Nick noted 25% of households pay for SiriusXM, 55% for cloud storage, and 91% for at least one streaming service.
  • The optimistic-but-blunter case — nobody actually uses it yet. Microsoft’s Nicholas Bustamante: “This is my thesis. No one uses AI. I repeat, absolutely no one. We live in a bubble. Even among my friends who pay for it, when I ask them to open ChatGPT and show me their queries, it’s the same handful of basic things. Most don’t even know they can upload a photo and ask questions about it. Connecting Gmail so an agent can read and send emails blows their minds. An agent opening a browser and checking them into a flight? They’ve never even heard of it.” He frames adoption — and getting existing subscribers to use what they pay for — as the massive challenge, and asks readers to imagine the compute shortage if everyone used AI the way today’s top 1% do.
  • The counterpoint — there is no utility here. Lawyer Basil Musharbash, reposting Bustamante: “I’m a competent adult. Why do I need an AI agent to check into a flight for me? I’m literate. Why do I need an AI agent to read and write emails for me? The reason no one uses AI is that there’s no serious utility here for the vast majority of people.”
  • The sharpest version of that objection targets the flight example. Habibi Slop: checking into a flight has never taken 20 minutes — “open the airline’s app, click the big red check-in button, click number of check bags, done” — and verifying the agent succeeded costs the same effort. Further, about 55% of American adults take zero flights in a given year, so few people care about the use case at all. Their conclusion: “People underestimate AI agents is not what’s happening here. What’s happening is tech people routinely mistake their own unusually high travel, high email, high admin lifestyle for a universal human workload.” They characterise the overarching problem of the last 20 years of tech as an inability to look outside itself, becoming obsessed with marketing solutions to problems nobody else has.
  • An even more pessimistic read of the consumer. Breaking Points hosts Krystal and Saagar (rendered in the transcript as “Cigar and Jetty”): “Nobody actually needs a personal AI assistant. The average screen time in the U.S. is seven hours a day. People have plenty of time. There’s no evidence that freeing up more will lead to anything but more TikTok scrolling. If anything, daily life needs more offline friction. ‘I’ll spend it doing what I love.’ No, you won’t. You’ll watch TikTok. The data doesn’t lie.”
  • The failure-of-imagination rebuttal. Investor Nick Carter, replying to Musharbash: “You have 150 IQ plus tireless genius living in your computer. It can use your browser and read your email and access any file you want. It won’t complain if you give it the most tedious administrative tasks. It’s willing to work 24-7 for less than minimum wage. It can work unsupervised for hours at a time. And you can’t think of any way to put it to use?” — listing paying parking tickets, triaging emails, manually unsubscribing from spam, vetting credit card payments, cancelling subscriptions, and renegotiating insurance.
  • The host’s own position splits the difference. He bristles at the idea that everyone is too lazy and slack-jawed to get off TikTok and use these tools, even if productivity isn’t their primary motivation. But he also observes that nearly everyone he sees using AI assertively and positively is a “go-getter” type — people who want more from life, jobs, and careers, not necessarily entrepreneurs but bringing entrepreneurial energy. They are, in short, high aspiration — and not everyone is.

Why a ceiling on average-consumer spend need not break the financial thesis

  • Power-user spend has far less of an upper bound than SaaS. Unlike previous software categories, where the ceiling was the most premium account at $100 or $200 a month, AI power users and the companies they work with can spend thousands, tens of thousands, per month in a way that remains ROI-positive.
  • Consumer AI need not be productivity software, nor subscription-priced. The host returns to Suno, whose hundreds of millions of dollars of ARR come not only from musicians using it as a productivity tool but from “a lot of normies” using it as an entertainment platform.
  • Advertising — the model “we love to hate,” but which funds most of the internet — is arriving on schedule. OpenAI announced a new in-line visual ad format for ChatGPT, beyond text ads: “Our new visual ad format helps people imagine how products and services could fit into their lives. Through images showing product inspiration, product usage, or the experiences they make possible, people can discover a new brand, understand what makes an offer distinctive, and decide what to buy. For advertisers, this creates new opportunities to tell their story through images.”
  • Scale and headroom: the same announcement reiterated a figure also given at Dev Day — ChatGPT now reaches 1.2 billion people each week, up 20% since the one-billion milestone announced at the end of July. Alex Immerman: ChatGPT “launched ads in February, crossed 100 million run rate revenue in six weeks, and a billion by August. Google and Meta combined for $500 billion in ad revenue. We are so early in AI ads.”

The “room to grow” matrix (chart described)

The host calls this the most interesting chart in the a16z report for anyone trying to understand consumer AI’s future. It is a three-column table:

Category                        | Web 1.0 / Web 2.0 winner(s) | AI company in top 100?
--------------------------------+-----------------------------+-----------------------
Productivity tools              |            ...              |  YES
Search & answers                |            ...              |  YES
Production & creation           |            ...              |  YES
Photos & video                  |            ...              |  YES
Education                       |            ...              |  some
Health                          |            ...              |  some
--------------------------------+-----------------------------+-----------------------
Streaming & media platforms     |            ...              |  NONE
Retail marketplaces             |            ...              |  NONE
Social networks & messaging     |            ...              |  NONE
Travel                          |            ...              |  NONE
Personal finance                |            ...              |  NONE
Jobs & career networks          |            ...              |  NONE
Real estate                     |            ...              |  NONE
Gaming                          |            ...              |  NONE
Dating                          |            ...              |  NONE

Left column: the category. Middle: which companies won it in Web 1.0 and Web 2.0. Right: whether any AI company currently sits in the consumer top 100. AI presence is highly concentrated in productivity, search and answers, production and creation, and photos and videos, with a little education and health — and entirely absent from the nine categories below the line.

The host closes by granting that it is sometimes too easy for early-adopter types to say “we’re so early” when looking at underwhelming adoption numbers — but says it would nonetheless be a mistake to assume that where consumer AI is today is where it will be a couple of years from now.

Key Concepts

  • Beam — Reflection AI’s 501-billion-parameter US-trained open-weight model, pitched on inference efficiency (3–4× GLM 5.2) as a cost-effective base for enterprise post-training.
  • Benchmark maxing — optimising a model for a specific published benchmark rather than general capability; the show cites Terminal Bench 4.0 as having exposed a lot of it, and notes Reflection’s omission of that benchmark.
  • Distributed RL infrastructure — running a reinforcement-learning post-training stage across distributed compute; described as a wildly underexplored method for training open models.
  • Agent fleet — a large group of agents running independently against the same target, deliberately distinguished from an “agent swarm” by the absence of evidence that the agents coordinate with each other.
  • Agent swarm chasing — the emergent research practice of detecting, attributing, and documenting large-scale autonomous agent activity on the internet.
  • Customer concentration — the share of revenue coming from a handful of customers; readable either as revenue risk or as evidence of early adoption with headroom in the rest of the market.
  • Top 100 Consumer AI Apps — Andreessen Horowitz’s recurring report, now in its seventh edition, ranking consumer AI by unique monthly web visits, mobile MAU, and (new this edition, via YipitData) monthly revenue.
  • Consolidation phase — the stage at which a new market stops producing many new entrants and begins exhibiting stable patterns and network effects within categories; inferred here from only 11 first-time debuts.
  • Power-user concentration — the finding that the top 1% of AI spenders outspend the bottom 50% combined ($903/month vs. a $25 median).
  • TAM (total addressable market) — the full revenue opportunity for a product; invoked by optimists to reframe the 98% of non-paying households as upside.
  • Adoption curve comparison — using the time earlier technologies took to go from ~2% to ~50% US adoption (computer 19 years, cars 15, smartphone 6.5) as a benchmark for AI’s trajectory.
  • High-aspiration user — the host’s term for the people he observes using AI assertively: not necessarily entrepreneurs, but bringing entrepreneurial energy and wanting more from their work and lives.
  • Visual in-line ads — OpenAI’s new ChatGPT ad format using images of product inspiration, usage, or resulting experiences, extending beyond text-based ads.
  • Room-to-grow matrix — the a16z chart mapping consumer categories against their Web 1.0/2.0 winners and whether any AI company yet ranks in the consumer top 100.
  • Plaud — an AI note-taking device whose walk-around form factor the host offers as an example of a popular product that escapes insider notice.
  • Suno — AI music generation platform at number seven on the revenue chart, cited as proof that consumer AI revenue can come from entertainment rather than productivity.
  • n8n — workflow automation platform, a precursor to the agentic era, for which the top 1% of spenders are ~23× as likely to pay as other buyers.
  • OpenRouter — a single-platform gateway to models from multiple providers; its rank at 13 on a consumer chart is read as evidence of demand for cost optimisation.

Summary

The episode uses the seventh edition of a16z’s Top 100 Consumer AI Apps report to argue that consumer AI is currently two radically different markets wearing one name. At one extreme, a tiny cohort of power users spends an average of $903 a month — the top 1% of payers now outspending the bottom 50% combined — overwhelmingly on tools to build, automate, create, and work. At the other, 98% of US households pay nothing at all, and the episode presents in full the case that many of them never will: that agentic use cases like flight check-in and email triage are tech workers projecting their own unusually admin-heavy lives onto a public that neither needs nor wants the help. The host declines to pick a winner outright, rejecting the premise that ordinary people are simply too lazy to benefit, while conceding that the people he sees using AI well are consistently high-aspiration types, and not everyone is. His resolution is that the question may be less decisive than it looks: unlike SaaS, AI has almost no ceiling on what a power user can profitably spend, and consumer AI need be neither productivity software nor subscription-funded — Suno’s hundreds of millions in ARR come substantially from entertainment, and OpenAI’s new visual in-line ads, atop 1.2 billion weekly users, point to the advertising model that already funds most of the internet. The closing note is a caution against both reflexes: insiders too readily excuse weak adoption with “we’re so early,” but it would still be a mistake to assume today’s consumer AI landscape — concentrated in productivity, search, creation, and media, and entirely absent from streaming, retail, social, travel, finance, jobs, real estate, gaming, and dating — is where it will stand in a couple of years.